Artificial Intelligence · arXiv:2611.02797 v1 · submitted 2 Nov 2026 · 36 references
SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents
The paper says
“Recursively self-improving (RSI) systems are increasingly viewed as a promising means of accelerating progress toward artificial general intelligence.”
more of the abstract
They aim to autonomously improve their capabilities by refining their working pipelines and recursively improving the strategies that guide subsequent refinements, with gains generalizing to unseen scenarios.However, existing evaluations lack reliable metrics to distinguish capability gains from scenario fitting, sustainable and extensible protocols for unseen tasks and difficulties, and fine-grained attribution of improvement to evolution traces. To this end, we propose SelfImprove-Bench,an RSI evaluation framework that organizes tasks into three primary capabilities (i.e., Exploring, Planning, Reasoning) as scenarios for guiding model evolution. We instantiated the tasks on games for controllability and reproducibility, and pairs them with scalable generation and orchestration driven by task templates and difficulty coefficients. Beyond task-level scores, SelfImprove-Bench organizes evaluation into generalization ladder levels G0 to G4 that measure how far gains transfer from training conditions. Experiments reveal that the evolutionary gains of existing RSI methods are inconsistent across generalization levels.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 3 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The benchmark design is a real contribution; the headline generalization-reversal numbers are not yet supported at the claimed precision. This line is wrong
4 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 36 references say.
The paper says, at Section 4.2, Table 2, Figure 4; Eq. (1)
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The headline claims in Section 4.2 are built from single HCI point estimates: Table 2 and Figure 4 report one HCI number per cell, computed by Eq. (1) from means, with no standard errors, confidence intervals, evaluation-set sizes, or statement of the number of independent refinement runs.
more of objection 1
Appendix A.1 defines the relevant fields (holdout_tasks_per_difficulty, count, seeds) but the realized values are never reported; Section 4.1 gives sizes only for MC and Astray (five fixed tasks each). The advertised reversals are therefore differences between unquantified estimates: the 109.2%-vs-100.0% Combat-to-Farm result (b111) is a 0.012 gap between mean scores 1.191 and 1.179 against a pseudo-ground-truth that the paper itself says can be exceeded (b268), and the reverse-direction Reasoning effects (b110) are -3.0, -2.2, and -1.7 HCI points. Because Eq. (1) divides by (sr - sb), HCI is unstable near sb = sr; the denominator already degenerates for AHE on MC at G2, producing the unexplained N/A in Table 2. Seed-level variation in instance scores or in refinement trajectories could plausibly flip these signs or close the gap. The paper should report per-cell instance counts, per-instance score distributions or standard errors, the number of runs per method, and bootstrap confidence intervals for HCI, then restate which reversal claims survive at the achieved precision.
The paper says, at Sections 3.1-3.2; G3/G4 results in Section 4.2
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The G3 and G4 conclusions presuppose the Section 3.2 capability mapping (Planning: Farm/Combat; Reasoning: Hero/Sweepmine; Exploring: MC/Astray), but the mapping is asserted rather than validated, and Section 3.1's claim that games avoid the confounding of multiple abilities is stronger than the design supports: Sweepmine requires planning the order of information-gathering, and MC and Astray require planning in service of exploration.
more of objection 2
The direction-dependent transfer claim (b110) and the exceeds-direct-refinement claims (b111) measure whatever this unvalidated taxonomy labels, and the G3 pairings are a small set of design choices that could drive the observed direction effects. The paper should supply construct-validity evidence, for example within-category versus cross-category correlations of A0 scores, capability probes, or trace-based evidence that the games differentially engage the three capabilities, or it should soften the capability-level interpretation of G3/G4.
The paper says, at Section 4.2, G2 paragraph
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The sentence on Farm, HyperAgents beats AHE at the training difficulty but trails it on the Hard held-out set is inconsistent with the metric the section uses. Under HCI, Table 2 shows HyperAgents ahead at both levels (G1: +100.0 vs -33.8; G2: +19.3 vs -1.8).
more of objection 3
The claim holds only for raw final scores (1.179 vs 1.080 at G1; 0.942 vs 1.069 at G2), and presenting a raw-score reversal inside an HCI-based discussion is misleading; moreover, under HCI the ranking reversal the paragraph needs is actually supplied by the Astray row (G1: +20.0 vs +6.7; G2: -2.9 vs +20.0). The authors should state explicitly which quantity reverses and correct the example so that text, table, and metric agree.
The paper says, at Section 4.2, G4 paragraph; Figure 4
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The G4 claims (b113: the Farm-refined agent on Sweepmine and Hero and the Combat-refined agent on Hero exceed agents refined directly on those targets) could not be verified against Figure 4 as rendered. The Sweepmine panel shows Target: +100.0% with source values -41.2 (Farm) and -58.0 (Hero), the Hero panel shows Target: +100.0% with -12.0 (Farm) and +17.3 (Combat), and …
more of objection 4
… the Combat panel shows Target: +12.9% with +21.2 (Farm) and +10.7 (Sweepmine); on their face these numbers support only the Farm-to-Combat case. The five-of-eight-directions count in the text is also not recoverable from the rendered figure. The figure and the G4 text need to be reconciled so the direction-level claims can be checked.
5 minor objections
The paper says, at Section 4.2 and Figure 1
Pith says: minor objection 1 of 5
There are typos in the manuscript: Tablee 2 in Section 4.2 should be Table 2, and Figure 1 contains Generaliaztion Comparision and Generaliztion; these should be corrected.
The paper says, at Table 2, MC G2 row
Pith says: minor objection 2 of 5
The N/A for AHE on MC at G2 should be footnoted: by Eq. (1) the denominator (sr - sb) vanishes when the baseline equals the reference, so the entry reflects a degenerate ratio rather than a missing experiment; readers cannot otherwise distinguish the two cases.
show 3 more
The paper says, at Appendix A.1, b209
Pith says: minor objection 3 of 5
The paper acknowledges that descriptor-level deduplication does not guarantee distinct underlying game states; reporting the frequency with which different descriptors map to identical or near-identical states in the released pools would let readers assess how novel the G1 held-out instances actually are.
The paper says, at Section 4.3 and Table 7
Pith says: minor objection 4 of 5
The annotation agreement is honestly labeled as post-review, but reporting the machine-only agreement before human correction would strengthen the trace-analysis claims; the current table cannot be read as a validation of the automatic annotation.
The paper says, at Section 4.1 and Table 2 caption
Pith says: minor objection 5 of 5
Because MC and Astray use only five fixed evaluation tasks, the MC-Astray G3/G4 comparisons rest on five target instances per direction; this should be restated in the Table 2 caption and factored into the strength of the cross-category claims.
What would settle it
Re-run the G3 transfer cells, especially HyperAgents Combat-to-Farm and all three methods' Hero-Sweepmine direction pairs, with at least 10 independent held-out seeds while keeping training sequences and refinement fixed; report per-seed HCI and a bootstrap confidence interval over target instances, along with the sr values used per game and difficulty. If the 109.2-versus-100.0 gap falls inside the interval or the sign of the Hero/Sweepmine direction effect flips across seeds, the ranking-reversal and direction-dependence claims are not established.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 3 Nov 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
SelfImprove-Bench is worth knowing about. The G0–G4 ladder — replay, held-out same difficulty, harder instances, within-category transfer, cross-category transfer — is a clean way to organize generalization for non-training RSI agents. The protocol is careful: held-out pool constructed first, descriptor-hash deduplication, feedback restricted to a defined interface, candidate validation on held-in tasks only. Table 1's claim that no cited benchmark combines these axes is credible. The trace analysis is coarser but still an addition. The authors are also honest about the pseudo-ground-truth being approximate (scores can exceed it, b268) and about annotation being post-review rather than blind (b232).
The soft spots are empirical. Every HCI in Table 2 is a point estimate, no error bars, and the paper never states evaluation sizes for the main runs; MC and Astray use five fixed tasks. The two most-quoted findings are fragile. HyperAgents' 109.2% from Combat to Farm versus 100.0% direct is a 0.012 score gap above a reference that is itself approximate. The Hero-to-Sweepmine 'degradation' is -3.0, -2.2, and -1.7 HCI points — all within a few points of zero. Since Eq. (1) divides by (sr - sb), HCI is unstable when the baseline is near the reference; the N/A for AHE on MC at G2 is a symptom. Seed-level variation could flip signs. There is also an internal inconsistency: the text says HyperAgents beats AHE at the training difficulty but trails on the Hard held-out set on Farm (Section 4.2), yet Table 2 shows HyperAgents at +19.3 and AHE at -1.8 on G2 Farm. The sentence contradicts its own table.
The G3/G4 interpretation rests on the asserted capability labels: Farm and Combat as Planning, Hero and Sweepmine as Reasoning, MC and Astray as Exploring. That categorization is a modeling judgment without independent validation, and the direction-dependent claims ride on it. If Sweepmine and Hero overlap in planning as well as reasoning, the 'Reasoning' label oversimplifies.
The circularity concern does not land: HCI and the RSI definition come from the companion paper [34], but no constant was tuned to produce the observed inconsistency. That is legitimate reuse.
The coarse conclusion — gains vary across the ladder — survives. The sharp reversals do not, at the claimed precision. This paper is for people building RSI evaluation harnesses; the ladder and generation protocol deserve adoption. Send it to peer review, not a desk reject, because the design is worth hardening with uncertainty measures, larger evaluation sets, and a corrected text. But the empirical conclusions should not be accepted as stated.
The strongest objection
Table 2 and Figure 4 report every G0-G4 result as one HCI number per cell, computed by Eq. (1) from means, with no standard errors, confidence intervals, or stated evaluation-set sizes; Appendix A.1 defines counts but the paper never gives them. The directional findings that carry the paper (b110, b111) are differences between these point estimates. Combat-to-Farm HyperAgents reaches 109.2% HCI because its average final score is 1.191 against a reference sr that the direct Farm-refined agent exactly reaches (1.179), a gap of 0.012 on an approximate pseudo-ground-truth that the paper itself says can be exceeded (b268). The three reverse-direction Reasoning degradations are -3.0, -2.2, and -1.7 HCI points, all within a few points of zero. Because Eq. (1) divides by (sr - sb), HCI is unstable or undefined when the baseline is near the reference; the paper already reports N/A for AHE on MC at G2. Seed-level variation in instance scores could therefore flip the sign of the direction effects or close the 109.2-versus-100.0 gap. The benchmark design may be sound, but the paper's headline empirical conclusion of inconsistent generalization is not yet supported at the precision claimed.
The reader's weakest assumption was the capability categorization in Section 3.2, and that is a genuine construct-validity concern: the G3/G4 contrast depends on Farm and Combat being Planning, Sweepmine and Hero being Reasoning, and MC and Astray being Exploring, with only one transfer pair per category. I agree with that concern, but the more load-bearing weakness is upstream: even if the labels are accepted, the empirical claims that make the paper interesting are comparisons of HCI point estimates without any uncertainty quantification. The reader mentioned small evaluation sets and missing error bars in the rationale but did not make this the weakest assumption. This matters because the specific reversal used as headline evidence is a 0.012 absolute score difference above an approximate reference, and the reverse-direction Reasoning effects are within a few HCI points of zero, so seed noise or reference error could change the conclusions. This does not invalidate the benchmark as a protocol, and it is fixable by reporting seed-level or bootstrap uncertainty, so CONDITIONAL remains the appropriate verdict and no adjustment is needed.
The referee's report
If the empirical claims survive added uncertainty quantification, this is a useful contribution. The protocol is unusually careful for a first benchmark: held-out construction precedes held-in generation; descriptor deduplication spans both pools; feedback is confined to a published interface; candidate validation cannot access S_out; resolved configurations and task manifests are saved for reconstruction; and the G0-G4 ladder gives the community a shared language for RSI generalization. The paper is also honest where it matters: it states that descriptor deduplication is not state-level (b209), that the Farm pseudo-ground-truth search can be beaten (b267-b268), and that the Table 7 annotation agreement is post-review rather than blind (b232). The schedule ablation (Table 4) and the HyperDSH study (Tables 5-6) are falsifiable design experiments that go beyond headline scores. The measured inconsistency of gains across the ladder, if confirmed with error bars, would be a substantive community-wide caution for how RSI results are reported.
Pith's reading of the paper (a machine's summary, not the authors')
This paper argues that a single score cannot tell you whether a self-improving AI agent has genuinely gotten better, and it builds a benchmark to demonstrate why. SelfImprove-Bench runs six games grouped into three capabilities — Farm and Combat as Planning, Hero and Sweepmine as Reasoning, MC and Astray as Exploring — and scores an agent at five distances from its training data: replaying seen instances, fresh instances at the same difficulty, harder instances, a new game in the same capability, and a game in a different capability. Evaluated on three representative non-training self-improving methods, the benchmark reports that gains are inconsistent across those distances: methods that improve on fresh instances can lose on harder ones, transfer within a capability works in one direction but not the reverse, and an agent refined on a sibling game can beat an agent refined on the target itself. If correct, similar final scores routinely conceal opposite generalization behavior, so claims about self-improving agents should be read level by level rather than from an endpoint score.
What carries the argument. The load-bearing object is the generalization ladder, five transfer distances defined on top of six template-generated games. G0 replays previously seen instances; G1 uses held-out instances at training difficulty; G2 raises difficulty within the same game; G3 moves to a different game inside the same capability category (Farm–Combat for Planning, Sweepmine–Hero for Reasoning, MC–Astray for Exploring); G4 moves to a game in a different capability. Gains at every level are normalized with the Headroom-Closed Index, $$HCI = \frac{s_a - s_b}{s_r - s_b}$$, which measures the fraction of the gap between a baseline score and a reference score that a refined agent closes, so that gains are comparable across games with very different raw scores. Two supporting mechanisms make the ladder usable: template- and difficulty-coefficient-based generation, which produces fresh held-out instances and keeps the task pool from saturating as agents improve, and semantic trace analysis, which classifies each refinement proposal by functional target, applicability scope, and modification surface so that a method's changes can be connected to where its gains do and do not transfer.
Core claim. The central claim is that a generalization ladder can separate genuine self-improvement from scenario fitting, and that on such a ladder the evolutionary gains of existing RSI methods are inconsistent. The empirical core is an evaluation of three representative non-training methods — ACE (experience-level), AHE (harness-level), and HyperAgents (meta-level) — across six games. Sixteen of eighteen method–game pairs improve on unseen instances at training difficulty (G1), yet gains fail to extend reliably to harder instances (G2); among the two Reasoning games, all three methods improve transferring from Hero to Sweepmine but degrade in the reverse direction (G3); and HyperAgents' cross-category transfer is positive in only five of eight directions (G4). A source-game agent can even beat direct refinement on the target, reaching 109.2% HCI from Combat to Farm versus 100.0% from refining on Farm itself. The paper concludes that similar final scores conceal systematically different evolution patterns and transfer ability, and that method rankings reverse depending on which generalization level is measured.
The premise it rests on. The ladder's G3-versus-G4 contrast stands or falls with the paper's classification of Farm and Combat as Planning, Hero and Sweepmine as Reasoning, and MC and Astray as Exploring — a modeling judgment with no independent validation — so if the games overlap capabilities, the observed transfer patterns would not measure what the ladder labels.
If the paper is right (Pith's inferences, not the authors')
Endpoint scores on held-out tasks are insufficient: evaluations of self-improving agents should report gains at several transfer distances, since the same method can look best at G1 and worst at G3.
A gain on one game does not license a claim about its capability: direction-dependent transfer means refinement success and failure must be reported per direction, as all three methods improved from Hero to Sweepmine but degraded in reverse.
Task scheduling choices can matter as much as method choices: random medium-difficulty training outperformed fixed-instance training on both held-out Combat sets and Hard Hero tasks, while harder or mixed-difficulty training did not consistently help.
A target's own refinement is not an upper bound on performance: an agent refined on a sibling game reached 109.2% HCI on Farm versus 100.0% from refining on Farm itself, so cross-task refinement can substitute for direct refinement.
Revision activity is a visible correlate of improvement: AHE and HyperAgents revise their recorded opinions more often than ACE (30.6% and 40.2% versus 14.2%) and generally finish stronger, suggesting effective self-improvement depends on revising strategies, not only on adding them.
The direction asymmetry reported for the Reasoning pair (improvement from Hero to Sweepmine, degradation in reverse) could be driven by asymmetric difficulty or headroom between the two games rather than by the capability label; an editor's test is to match Sweepmine and Hero baseline distributions and re-run the G3 protocol.
The G0–G4 ladder is a transferable template: any domain with a template-based instance generator, such as software engineering or web-agent benchmarks, could layer the same replay/instance/difficulty/task/capability hierarchy onto itself and gain the same separation of fitting from generalization.
The result that a source-game agent can beat direct refinement on the target suggests a curriculum implication the paper leaves implicit: chaining refinement across a sequence of games might produce a stronger target agent than refining on the target alone, and the scheduling result (random medium-difficulty training beating fixed-instance training) points in the same direction.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Recursively self-improving (RSI) systems are increasingly viewed as a promising means of accelerating progress toward artificial general intelligence. They aim to autonomously improve their capabilities by refining their working pipelines and recursively improving the strategies that guide subsequent refinements, with gains generalizing to unseen scenarios.However, existing evaluations lack reliable metrics to distinguish capability gains from scenario fitting, sustainable and extensible protocols for unseen tasks and difficulties, and fine-grained attribution of improvement to evolution traces. To this end, we propose SelfImprove-Bench,an RSI evaluation framework that organizes tasks into three primary capabilities (i.e., Exploring, Planning, Reasoning) as scenarios for guiding model evolution. We instantiated the tasks on games for controllability and reproducibility, and pairs them with scalable generation and orchestration driven by task templates and difficulty coefficients. Beyond task-level scores, SelfImprove-Bench organizes evaluation into generalization ladder levels G0 to G4 that measure how far gains transfer from training conditions. Experiments reveal that the evolutionary gains of existing RSI methods are inconsistent across generalization levels.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-03
The authors' main claim, as the classifier picked it out 1.00
“Experiments reveal that the evolutionary gains of existing RSI methods are inconsistent across generalization levels.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.18 · human participants 0.04 · controlled comparison 0.13 · randomized assignment 0.02 · states quantitative result 0.11 · compares with prior results 0.64 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Recursively self-improving (RSI) systems are increasingly viewed as a promising means of accelerating progress toward artificial general intelligence. | background 1.00 | 0.04 | 0.03 | asserted 0.59 |
| They aim to autonomously improve their capabilities by refining their working pipelines and recursively improving the strategies that guide subsequent refinements, with gains generalizing to unseen scenarios.However, existing evaluations lack reliable metrics to distinguish capability gains from scenario fitting, sustainable and extensible protocols for unseen tasks and difficulties, and fine-grained attribution of improvement to evolution traces. | gap 0.97 | 0.07 | 0.07 | asserted 0.99 |
| To this end, we propose SelfImprove-Bench,an RSI evaluation framework that organizes tasks into three primary capabilities (i.e., Exploring, Planning, Reasoning) as scenarios for guiding model evolution. | contribution 0.89 | 0.94 | 0.05 | asserted 1.00 |
| We instantiated the tasks on games for controllability and reproducibility, and pairs them with scalable generation and orchestration driven by task templates and difficulty coefficients. | method 0.99 | 0.93 | 0.03 | asserted 1.00 |
| Beyond task-level scores, SelfImprove-Bench organizes evaluation into generalization ladder levels G0 to G4 that measure how far gains transfer from training conditions. | method 0.84 | 0.87 | 0.07 | asserted 1.00 |
| Experiments reveal that the evolutionary gains of existing RSI methods are inconsistent across generalization levels. | result 1.00 | 0.93 | 0.64 | asserted 1.00 |
What kind of paper the classifier reports this is: dataset or benchmark 0.99 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-03: 1 request, 5,054 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
References (36)
- [34] Y. Duan, Y. Liu, Z. Tang, H. Chen, J. Zhou, Y. Liu, B. Xu, Y. Wu, S. Chen, Y. Zhou, H. Wang, X. Yu, S. Han, X. Zhu, L. Zhou, B. Lu, W. Zhou, J. Liu, N. Fang, J. Tian, R. Chen, Y. Li, K. Zuo, K. Zhang, Q. Yang, Z. Wang, J. Qiu, C. He, G. Li, B. Zhou, Z. Liu, Z. Wen, J. Kang, X.... on PithPith's reviewarXiv
- [1] J. Zhang, B. Zhao, W. Yang, J. Foerster, J. Clune, M. Jiang, S. Devlin, and T. Shavrina, “Hyperagents,” arXiv preprint arXiv:2603.19461, 2026. [Online]. Available: https://arxiv.org/abs/2603.19461 on PithPith's reviewarXiv
- [2] Q. Zhang, C. Hu, S. Upasani, B. Ma, F. Hong, V. Kamanuru, J. Rainton, C. Wu, M. Ji, H. Li, U. Thakker, J. Zou, and K. Olukotun, “Agentic context engineering: Evolving contexts for self-improving language models,” arXiv preprint arXiv:2510.04618, 2025. [Online]. Available: http... on PithPith's reviewarXiv
- [3] J. Lin, S. Liu, C. Pan, L. Lin, S. Dou, Z. Xi, X. Huang, H. Yan, Z. Han, T. Gui, and Y.-G. Jiang, “Agentic harness engineering: Observability-driven automatic evolution of coding-agent harnesses,”arXiv preprint arXiv:2604.25850, 2026. [Online]. Available: https://arxiv.org/abs... on PithPith's reviewarXiv
- [4] C. Zheng, C. Xue, B. Liang, J. Yang, and C. Zhang, “Seagym: An evaluation environment for self-evolving llm agents,”arXiv preprint arXiv:2606.17546, 2026 on PithPith's reviewarXiv
- [5] L. Huang, C. Yang, H. Zhou, H. Song, Z. Chen, R. Le, Y. Song, W. X. Zhao, and T. Zhang, “Evo-Bench: Can language models improve agent harness?” arXiv preprint arXiv:2608.09096, 2026. [Online]. Available: https://arxiv.org/abs/2608.09096v1 on PithPith's reviewarXiv
- [6] B. Deng, K. Zhou, L. Guo, C. Tao, X. Chen, C. Xie, R. Liang, F. Chen, and C. Zhang, “FinEvo-Bench: A longitudinal benchmark for self-evolving agents in professional financial workflows,”arXiv preprint arXiv:2608.06144, 2026. [Online]. Available: https://arxiv.org/abs/2608.06144 on PithPith's reviewarXiv
- [7] Z. Stojanovski, O. Stanley, J. Sharratt, R. Jones, A. Adefioye, J. Kaddour, and A. Köpf, “Reasoning Gym: Reasoning environments for reinforcement learn- ing with verifiable rewards,” in Advances in Neural Information Processing Systems, vol. 38, 2025. [Online]. Available: http... on Pith
- [8] S. Xue, Z. Ding, Y. Shen, Y. Wang, Z. Yin, Y. Wu, Y. Chen, M. Wang, and L. Yang, “Past- bench: Benchmarking the foundations of recursive self-improvement in personal agents,”arXiv preprint arXiv:2608.04003, 2026 on PithPith's reviewarXiv
- [9] X. Gao, C. Hu, H. Chen, P. Yao, Z. Wang, Y. Bai, Z. Wu, Y. Han, X. Cong, J. Gui, Y. Deng, and T. Li, “EvoAgentBench: Benchmarking agent self-evolution via ability transfer,”arXiv preprint arXiv:2607.05202, 2026. [Online]. Available: https://arxiv.org/abs/2607.05202 on PithPith's reviewarXiv
- [10] S. Zhong, Y. Lu, J. Ning, Y. Wan, L. Feng, Y. Ao, L. F. R. Ribeiro, M. Dreyer, S. Ammirati, and C. Xiong, “SkillLearnBench: Benchmarking continual learning methods for agent skill generation on real-world tasks,” arXiv preprint arXiv:2604.20087, 2026. [Online]. Available: http... on PithPith's reviewarXiv
- [11] J. Zheng, X. Cai, Q. Li, D. Zhang, Z. Li, Y. Zhang, L. Song, and Q. Ma, “LifelongAgentBench: Evaluating LLM agents as lifelong learners,”arXiv preprint arXiv:2505.11942, 2025. [Online]. Available: https://arxiv.org/abs/2505.11942 on PithPith's reviewarXiv
- [12] H. Lin, C. Liu, X. Bai, X. Jin, Y. Li, N. Zheng, and X. Cao, “Beyond endpoint performance: Process-level evaluation of self-evolving agents,” arXiv preprint arXiv:2609.24663, 2026. [Online]. Available: https://arxiv.org/abs/2609.24663 on PithPith's reviewarXiv
- [13] H. Zheng, Y. Di, R. Sadhukhan, S. Jin, and B. Chen, “ServeLearnBench: How well can agents self-improve from serving experience?” arXiv preprint arXiv:2610.07792, 2026. [Online]. Available: https://arxiv.org/abs/2610.07792 on PithPith's reviewarXiv
- [14] E. Zelikman, E. Lorch, L. Mackey, and A. T. Kalai, “Self-taught optimizer (STOP): Recursively self-improving code generation,” inFirst Conference on Language Modeling, 2024. [Online]. Available: https://arxiv.org/abs/2310.02304v3 12 SelfImprove-Bench on PithPith's reviewarXiv
- [15] S. Hu, C. Lu, and J. Clune, “Automated design of agentic systems,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://proceedings.iclr.cc/paper_ files/paper/2025/hash/36b7acf6f6010652b3f2a433774a66fe-Abstract-Conference.html on Pith
- [16] P. Xia, R. Han, Z. Wang, Y. Chen, Y. Zhang, Y. Lee, C. Huang, H. Yu, Z. CuiZhu, Y. Ming, H. Yao, B. Gokturk, T. Pfister, and C.-Y. Lee, “RRSI: Regularized recursive self-improvement of agent harnesses,” arXiv preprint arXiv:2609.24972, 2026. [Online]. Available: https://arxiv.... on PithPith's reviewarXiv
- [17] Y. Chi, W. Li, D. Hong, X. Wang, M. Gao, K. Yang, B. He, Y. Zheng, C. Xiao, and Q. Na, “Ai4ai- bench: Benchmarking llm agents in algorithmic design for recursive self-improvement,”arXiv preprint arXiv:2608.20318, 2026 on PithPith's reviewarXiv
- [18] F. Meng, L. Du, Q. Chen, Z. Zhao, H. Lu, M. Hu, and M. Q. Shieh, “RSIBench-Data: Benchmarking data-centric research for recursive self-improvement,”arXiv preprint arXiv:2607.25886, 2026. [Online]. Available: https://arxiv.org/abs/2607.25886 on PithPith's reviewarXiv
- [19] N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,”arXiv preprint arXiv:2303.11366, 2023. [Online]. Available: https://arxiv.org/abs/2303.11366 on PithPith's reviewarXiv
show 16 more
- [20] W. Xu, Z. Liang, K. Mei, H. Gao, J. Tan, and Y. Zhang, “A-MEM: Agentic memory for LLM agents,” arXiv preprint arXiv:2502.12110, 2025. [Online]. Available: https://arxiv.org/abs/2502.12110 on PithPith's reviewarXiv
- [21] M. Robeyns, M. Szummer, and L. Aitchison, “A self-improving coding agent,” arXiv preprint arXiv:2504.15228, 2025 on PithPith's reviewarXiv
- [22] H. Zhang, S. Zhang, K. Li, C. Zhang, Y. Chen, Y. Zhang, L. Bai, and S. Hu, “Self-harness: Harnesses that improve themselves,” arXiv preprint arXiv:2606.09498, 2026. [Online]. Available: https://arxiv.org/abs/2606.09498 on PithPith's reviewarXiv
- [23] J. Zhang, S. Hu, C. Lu, R. Lange, and J. Clune, “Darwin gödel machine: Open-ended evolution of self-improving agents,” inInternational Conference on Learning Representations, vol. 2026, 2026, pp. 104223–104294 on Pith
- [24] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instructions... on PithPith's reviewarXiv
- [25] E. Zelikman, Y. Wu, J. Mu, and N. D. Goodman, “STaR: Bootstrapping reasoning with reasoning,” arXiv preprint arXiv:2203.14465, 2022. [Online]. Available: https://arxiv.org/abs/2203.14465 on PithPith's reviewarXiv
- [26] E. Nivel, K. R. Thórisson, B. R. Steunebrink, H. Dindo, G. Pezzulo, M. Rodriguez, C. Hernandez, D. Ognibene, J. Schmidhuber, R. Sanz, H. P. Helgason, A. Chella, and G. K. Jonsson, “Bounded recursive self-improvement,” arXiv preprint arXiv:1312.6764, 2013. [Online]. Available: ... on PithPith's reviewarXiv
- [27] Y. Wang, Z. Zhang, M. Chi, K. Yu, Y. Li, M. Peng, B. Tong, C. Zhang, Y. Zhou, and J. Li, “EvoMemBench: Benchmarking agent memory from a self-evolving perspective,” arXiv preprint arXiv:2605.18421, 2026. [Online]. Available: https://arxiv.org/abs/2605.18421v2 on PithPith's reviewarXiv
- [28] T. Guan, Y. Wang, H. Yang, S. Cao, S. Liu, Y. Hu, J. Li, and M. Zhang, “ContinualSkillBench: Can LLM agents truly evolve their capabilities?”arXiv preprint arXiv:2608.03874, 2026. [Online]. Available: https://arxiv.org/abs/2608.03874 on PithPith's reviewarXiv
- [29] P. Asawa, C. M. Glaze, G. Orlanski, R. Ramakrishnan, B. Xu, A. Biswal, V. S. Chen, F. Sala, M. Zaharia, and J. E. Gonzalez, “Continual Learning Bench: Evaluating frontier AI systems in real-world stateful environments,” arXiv preprint arXiv:2606.05661, 2026. [Online]. Availabl... on PithPith's reviewarXiv
- [30] C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan, “SWE-bench: Can language models resolve real-world GitHub issues?”arXiv preprint arXiv:2310.06770, 2023. [Online]. Available: https://arxiv.org/abs/2310.06770 on PithPith's reviewarXiv
- [31] M. A. Merrill, A. G. Shaw, N. Carlini, B. Li, H. Raj, I. Bercovich, L. Shi, J. Y. Shin, T. Walshe, E. K. Buchananet al., “Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces,” arXiv preprint arXiv:2601.11868, 2026. [Online]. Available: https... on PithPith's reviewarXiv
- [32] S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y. Bisk, D. Fried, U. Alon, and G. Neubig, “WebArena: A realistic web environment for building autonomous agents,”arXiv preprint arXiv:2307.13854, 2023. [Online]. Available: https://arxiv.org/abs/2307.13854 on PithPith's reviewarXiv
- [33] L. Phan, A. Gatti, Z. Han, N. Li, J. Hu, H. Zhang, C. B. C. Zhang, M. Shaaban, J. Ling, S. Shi et al., “Humanity’s last exam,” arXiv preprint arXiv:2501.14249, 2025. [Online]. Available: https://arxiv.org/abs/2501.14249 on PithPith's reviewarXiv
- [35] M. Ouyang, S. Hu, K. Q. Lin, H. T. Ng, and M. Z. Shou, “GameWorld: Towards standardized and verifiable evaluation of multimodal game agents,” arXiv preprint arXiv:2604.07429, 2026. [Online]. Available: https://arxiv.org/abs/2604.07429 on PithPith's reviewarXiv
- [36] A. A. Zheng, K. Misra, D. I. Beaver, and J. J. Li, “Hero’s journey: Testing complex rule induction with text games,”arXiv preprint arXiv:2606.02556, 2026. A Evaluation Protocol Details A.1 Task Configurations The task-related fields ofθ determine which instances are generated ... on PithPith's reviewarXiv
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment